SOC 2 for AI agents: the audit-trail question

If you sell an AI agent into enterprise, the security review will ask a version of: “What did the agent do, and can you prove it?” Here's why, and what SOC 2 expects.

Why SOC 2 cares about agent activity

SOC 2 is about controls over security, availability, processing integrity, confidentiality and privacy. Reviewers commonly look at how you monitor and record system activity (the “common criteria” / CC series). When your system is an agent acting on a customer's behalf, “system activity” includes the agent's actions — so reviewers want to see that those actions are recorded and controlled.

Why a debug log usually isn't enough

• It's typically mutable — not tamper-evident, so it isn't strong evidence.
• It's not at the action + approval level — it captures stack traces, not “the agent issued this refund; a human approved it.”
• It often has no retention policy or exportable, reviewer-friendly format.

What to prepare

1. A record of agent actions and approvals (not just errors).
2. Made append-only and tamper-evident, with a retention policy.
3. Exportable as evidence when an auditor or customer asks.